import speed optimization

[dskarbrevik]

While work has been done to minimize the dependency list of uqlm as much as possible, some dependencies require the loading or downloading of large files. This can make the import of uqlm a time intense process.

This issue hopes to speed up import times by:

1. speed profiling imports to identify import bottlenecks
2. remove bottlenecks where possible by either restructuring code so these bottlenecks only run when absolutely needed or refactoring code so the bottlenecks simply don't exist

If after speed profiling we find that step 2 does not have any reasonable quick wins, we will at least know where the limitations lie and can prioritize improvements as future work.

[mehtajinesh]

@dylanbouchard Working on this.

[mehtajinesh]

@dskarbrevik @dylanbouchard @mohitcek

This is the output of the profiler for all the libs/dependencies used by uqlm:

• langchain: 0.0332 seconds
• langchain_core: 0.0000 seconds
• transformers: 1.2427 seconds
• sentence_transformers: 4.8847 seconds
• bert_score: 0.0143 seconds
• matplotlib: 0.0000 seconds
• optuna: 0.0244 seconds
• numpy: 0.0000 seconds
• pandas: 0.0000 seconds
• sklearn: 0.0000 seconds

This is the usecase where the import time goes beyond 5 secs because of sentence_transformers

• uqlm (with BlackBoxUQ usage): 5.1340 seconds

[mehtajinesh]

We can resolve this problem by looking at alternatives to sentence_transformers or smaller models to improve import speed. Is there an ideal import time we are targeting?